hypnguyen1209/offensive-claude

What it solves

This project provides a structured, spec-driven framework for performing offensive security research and penetration testing using Claude Code. It transforms a general-purpose AI assistant into a disciplined security researcher by enforcing a rigorous engagement workflow based on the Cyber Kill Chain, ensuring that actions are authorized, findings are evidence-backed, and the process follows industry-standard security phases.

How it works

The framework operates as a Claude Code plugin that injects a "SessionStart dispatcher" into every conversation. This dispatcher ensures that the AI follows a specific "discipline layer" (e.g., scope, OPSEC, and finding validation) before executing domain-specific skills.

Key components include:

• Engagement Workflow: A 9-phase pipeline (Scope $\rightarrow$ Recon $\rightarrow$ Weaponize $\rightarrow$ Delivery $\rightarrow$ Exploit $\rightarrow$ Installation $\rightarrow$ C2 $\rightarrow$ Actions $\rightarrow$ Report) with quality gates to prevent premature phase transitions.
• Skill Modules: 32 specialized domain skills (e.g., web pentesting, cloud security, AI red-teaming) that provide the AI with technique maps, reference libraries, and runnable scripts.
• Safety Guards: Executable Python scripts (like scope_guard.py and action_guard.py) that enforce target authorization and redact sensitive headers to prevent unauthorized or unsafe actions.
• Crash-to-Exploit Pipeline: A specialized toolchain using rr, gcov, and Z3 to turn software crashes into verified exploitability verdicts.
• Collaborative Agents: Eight specialized sub-agents (e.g., redteam-planner, finding-validator) that handle planning, execution, and adversarial validation of findings.

Who it’s for

It is designed for offensive security researchers, penetration testers, and red teamers who use Claude Code and want a systematic, reproducible, and safety-gated approach to vulnerability research and adversary simulation.

Highlights

• Spec-Driven Discipline: Enforces an "Iron Law" for security engagements, requiring authorization and proof before marking findings as confirmed.
• Comprehensive Skillset: Covers 32 distinct domains, including emerging fields like AI agent red-teaming and container escape.
• Automated Quality Gates: Validates that required artifacts and evidence (CWE, CVSS, ATT&CK IDs) exist before moving to the next phase of an engagement.
• Adversarial Validation: Uses a bounded generator-checker rebuttal loop to ensure findings are not hallucinations but are grounded in evidence.
• Integrated Toolchain: Includes a .devcontainer with binary analysis tools for deterministic root-cause analysis of memory corruption bugs.